Trend report · hn_ai · 2026-06-13
The five pillars of the post-AI interview
The conversation around AI-generated content has shifted. A year ago, the question was "can AI detect AI?" Now it's "what exactly is the stack looking for, and how do I stay under the radar?" That shift matters for anyone creating, publishing, or distributing content at scale — whether you're a creator dodging algorithm penalties or a recruiter trying to verify that the person you're hiring actually exists outside of a generative model.
What Platforms Actually Scan For in 2026
Modern AI-content detection isn't a single check. It's a layered pipeline that inspects your file at multiple levels. Here's what each layer looks for:
C2PA (Content Provenance) metadata. The Coalition for Content Provenance and Authenticity standard embeds cryptographically signed statements about a file's origin directly into the file container. Fields like c2pa.actions, c2pa.assertions, and c2pa.hashed_jumbf record whether a tool like Midjourney, Sora, or DALL-E created or modified the content. Major platforms including Adobe, Microsoft, and the BBC have adopted C2PA. When a JPEG or MP4 carries these blocks, detection is near-certain.
AI-specific metadata beyond C2PA. Before C2PA became standard, tools already left traces in EXIF, XMP, and container-level metadata. Software, Make, Model, and Generator fields in EXIF headers still flag images from Stable Diffusion, Flux, and similar pipelines. TikTok and Instagram parse these fields during upload, even when the UI doesn't surface any warning.
Encoder signatures. Each generative model produces artifacts in the pixel or audio domain that statistical classifiers can identify. These aren't metadata — they're embedded in the actual data. Models like Stable Diffusion tend to leave consistent noise patterns in high-frequency regions. Video generation models produce distinctive temporal artifacts in frame-to-frame transitions. Platforms run these through classifier heads trained on known outputs. The key field here is the EncodingProcess value in EXIF, which records the specific codec chain used — a signal that detection models weight heavily.
Missing or inconsistent GPS/EXIF provenance. Organic human content almost always carries some EXIF context: a GPS coordinate, a timestamp with realistic drift, a device model. AI-generated content often has no GPS data, or has GPS data that doesn't match the claimed device. Platforms treat absence of expected fields as a weak signal — it gets combined with other factors rather than used alone.
What Gets Flagged on Instagram and TikTok
Instagram's detection system runs in two stages. First, at upload, metadata and C2PA blocks are parsed. If c2pa.claim_generator matches a known AI tool (the field will read something like Midjourney/1.0 or Adobe Firefly/3.0), the content enters a review queue. Second, a classifier head evaluates the actual pixel data for known generative artifacts. Content that passes both stages still gets demoted in recommendation feeds — Instagram doesn't ban AI content outright, but it suppresses reach for content that fails provenance checks.
TikTok is more aggressive. Its detection pipeline checks for EncoderID values in the file header that correspond to AI video pipelines. If you upload a video generated with Runway Gen-3 or Kling AI, the encoder signature in the container header will include markers TikTok's classifiers are trained to recognize. The result is either a label ("AI-generated") pinned to the post, or in repeated cases, a reach restriction. TikTok also cross-references the upload device against its device fingerprint database — this is where phone identity becomes relevant.
Why Stripping Alone Isn't Enough
You might think: strip the metadata, remove C2PA blocks, and you're done. The problem is device identity. Both Instagram and TikTok maintain device fingerprint databases. They track not just what you upload but how you upload it — the device model, the carrier, the IP geolocation cluster, the upload client string. If you're uploading from a device that's been flagged for repeated AI content uploads, stripping metadata doesn't help. The platform knows from the transport layer, not the file layer.
This is why naive stripping fails: you're solving a file-level problem while the platform is solving a device-level one. The detection stack looks at your file's provenance and your upload context simultaneously. Remove one signal and the other carries the flag.
The Durable Fix: Strip + Clean Phone Identity
The only approach that reliably clears both signals is a two-step process:
- Strip all AI provenance from the file. Remove C2PA blocks, clear EXIF/XMP metadata, normalize
EncodingProcessfields, and strip any container-level encoder signatures. This eliminates the file-layer signal. - Inject clean phone identity at upload time. Route the upload through a fresh device context — new device model, new carrier profile, new IP cluster. This replaces the device-layer signal with one that reads as organic, human, and unconnected to prior AI uploads.
Step one handles the metadata and C2PA checks. Step two handles the device fingerprint and transport-layer inspection. Together, they address the full detection stack rather than just one layer of it.
For creators publishing on a schedule, this isn't a one-time fix — it's a workflow. Each upload needs a clean context. That means fresh device profiles, rotated carrier identities, and IP geolocations that match realistic human distribution patterns. The goal is to make each upload look like it came from a different person with a different phone, on a different network, in a different place.
What This Means for Post-AI Content Strategy
The five pillars of post-AI content aren't just about detection avoidance — they're about operating in a world where platforms have built permanent infrastructure to distinguish AI-generated content from human-generated content. That infrastructure is getting more accurate with every model release, because every new AI tool adds new training data to the classifiers.
The durable play isn't to hide AI content indefinitely. It's to understand the detection stack thoroughly enough to make strategic decisions about when provenance matters and when it doesn't. For high-stakes professional use — hiring, publishing, brand presence — clean device identity combined with stripped metadata is the baseline. Without it, you're relying on detection systems that are actively improving.
If you're managing content at scale and want to test the approach without financial friction, here's the starting point.
→ Try Calabi free at calabilabs.com — 10 cleans, no card.